360 Secure SkillHub
360 安全技能中心

Hotel Social Automator

该 Skill 通过 MCP 工具调用营销场景列表、内容生成、草稿轮询、反馈改写、账号查询与发布接口,完成酒店社交媒体文案与配图的生成、审核和立即发布或定时发布,并可选用飞书发送通知;运行需双令牌认证、积分额度、mcporter 与 curl,以及指定工作区目录的写入权限。

tourmind
tourmind
热度 36.95
最近更新 2026-05-11 17:43:14
分类:营销依赖:MCP 服务 / 第三方 API / 代码执行 / 读写文件应用场景:社交媒体 / 文案创作 / 日程安排 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add hotel-social-automator
Metadata
name
media.hotelmind
description
Generate, review, and publish social media content through MCP with AgentAuth and workspace token authentication.

media.hotelmind

Calls MCP tools to generate, review, and publish social media content.


Prerequisites

Required Environment

Dependency Purpose Required
mcporter CLI Invoke MCP server tools Yes
curl Download images from URLs Yes
feishu-send at /usr/local/bin/feishu-send Send Feishu notifications No (only if Feishu notifications needed)
Write access to /root/.openclaw/workspace/ Store temporary images Yes

Note: If feishu-send is not available, the skill will still function but cannot send notifications via Feishu.

Required Configuration

This skill requires dual-token MCP authentication:

  • Authorization: Bearer hp_sk_*
  • X-Agent-User-Key: uk_*

Users must:

  1. Sign in to AgentAuth and copy a uk_*
  2. Sign in to HotelPost and create or copy a workspace hp_sk_*
  3. Put both tokens into the installed agent's MCP configuration

Full user onboarding guide:

  • references/user-onboarding.md

MCP server must be configured in your agent's MCP settings:

{
  "mcpServers": {
    "hotelpost": {
      "type": "streamable-http",
      "url": "https://mcp.example.com/api/mcp",
      "headers": {
        "Authorization": "Bearer <API_KEY>",
        "X-Agent-User-Key": "<AGENTAUTH_USER_KEY>"
      }
    }
  }
}

Required credentials:

  • API Key: Obtain from HotelPost Web admin panel → Settings → API Keys. Format: hp_sk_xxxxxxxx
  • AgentAuth User Key: Obtain from AgentAuth Dashboard after user login. Format: uk_xxxxxxxx
  • MCP now requires both tokens together: workspace hp_sk_* plus user uk_*

Credit System

Content generation consumes credits:

  • Each draft generation (up to 4 candidates) costs 40 credits
  • Each regeneration costs 10 credits

If credit quota is exceeded, the tool returns AIQuotaExceededError. Ensure sufficient credits before generating content. Credits can be managed in the HotelPost Web admin panel.

Multi-Tenant Isolation

This skill uses the HotelPost API key to resolve the target workspace and uses X-Agent-User-Key to verify the calling user. Both are required, and all operations remain scoped to the workspace bound to the hp_sk_* key.


Tool Invocation

Must use mcporter to call MCP server tools — do NOT use exec/curl!

Correct usage:

mcporter call hotelpost.<tool_name> <parameters>

Examples:

  • mcporter call hotelpost.list_scenarios — List marketing scenarios
  • mcporter call hotelpost.generate_content scenarioId=scenario_demo_001 — Generate content
  • mcporter call hotelpost.get_drafts draftId=draft_demo_001 — Get draft details

Forbidden: using exec/curl to access MCP API directly. Must use mcporter!

MCP Server Configuration

MCP server name is hotelpost. Server URL and both auth headers must be configured as described above.

For full user-side setup, token acquisition, and troubleshooting:

  • references/user-onboarding.md

When to Activate

  • User asks to generate social media posts or publish content
  • User asks to browse or select marketing scenarios
  • User asks to manage, view, or modify drafts
  • User asks to check publishing status or post list
  • User mentions HotelPost, hotel marketing, or social media publishing
  • User asks to schedule posts
  • Keywords: hotel, social media, post, publish, draft, marketing scenario, Instagram, X, Facebook, LinkedIn

Core Flow

list_scenarios → generate_content → get_drafts (polling) → [regenerate_draft] → publish_post / schedule_post

Step-by-Step

  1. Select Scenario — Call list_scenarios, display available scenarios, let user choose
  2. Generate Content — Call generate_content(scenarioId, language), submit async generation task
  3. Poll Drafts — Call get_drafts(scenarioId), wait for status to change from GENERATING to ACTIVE (usually 15-30 seconds)
  4. View/Modify — Call get_drafts(draftId) to view full content and images; use regenerate_draft(draftId, feedback) to modify
  5. Publish — Call list_connections first to confirm available accounts, then call publish_post or schedule_post
  6. Check Status — Call get_post_status(draftId) to confirm publishing result

Tool Reference

Tool Purpose Key Parameters
list_scenarios List marketing scenarios none
generate_content Generate draft based on scenario scenarioId, language?
get_drafts Query draft status/details scenarioId or draftId
regenerate_draft Regenerate with feedback, supports language/platform draftId, feedback?, language?, platform?
list_connections List bound social media accounts none
publish_post Publish immediately draftId, connectionIds[]
schedule_post Schedule for later draftId, connectionIds[], scheduledAt
list_posts List posts status?, page?, pageSize?
get_post_status Get publishing status postId or draftId

regenerate_draft Language Options

  • zh — Chinese (default)
  • en — English
  • th — Thai
  • ms — Malay
  • ja — Japanese

regenerate_draft Platform Options

  • x — X (Twitter)
  • facebook — Facebook
  • instagram — Instagram
  • linkedin — LinkedIn

list_posts Status Options

  • SCHEDULED — Scheduled posts
  • PUBLISHING — Currently publishing
  • PUBLISHED — Successfully published
  • PARTIAL_SUCCESS — Partially succeeded
  • FAILED — Failed

Important Rules

Information Display

  • Do NOT expose internal info to users: Cards should NOT contain draft IDs, internal numbers, etc. Only show content and images the user needs to see.

Performance Optimization

  • Image URL optimization: The first get_drafts(scenarioId) response includes image URLs in the draft list. Do NOT call get_drafts(draftId) again to fetch URLs — use the URLs directly from the first call to save time.
  • Parallel processing:
    • Download images in parallel with & (curl -o "1.png" url1 & curl -o "2.png" url2 & wait)
    • Send multiple images in parallel (with rate limiting, 3-5 at a time recommended)
    • Avoid serial operations; parallelize whenever possible

Send Order

  • Send content first, wait, then send images. Do not send all content at once followed by images — users cannot correlate them.

General Rules

  • Must select a scenario before generating. Do NOT skip list_scenarios.
  • Generation is async. After generate_content returns, you MUST poll get_drafts at 10-15 second intervals, up to 3 times.
  • Drafts include image assets. No need for users to provide images.
  • Before publishing, must confirm account by calling list_connections to get connectionId.
  • If user says "post to Instagram", find INSTAGRAM platform ID from connections.
  • language defaults to zh (Chinese). Pass en when user requests English.
  • If user is not satisfied with draft, use regenerate_draft with feedback.
  • regenerate_draft supports specifying language (zh/en/th/ms/ja) and platform (x/facebook/instagram/linkedin). If not specified, regenerates for all platforms.

Image Handling

  1. Ensure /root/.openclaw/workspace/ directory exists before downloading
  2. Download images with random filenames like hotelpost_xxx.png
    • Do NOT save to /tmp/ — some platforms don't support it
  3. Image URLs must include full signature parameters! Truncated URLs will be inaccessible.
  4. Clean up temporary files after sending.

Workspace Directory

Before downloading images, ensure the workspace directory exists:

mkdir -p /root/.openclaw/workspace/

Platform Publishing

After publishing, send notifications via the appropriate IM platform. See the platform-specific guide in references/:

  • feishu-guide.md — Feishu platform sending guide (card/image format, ID selection rules)

Note: Feishu notification is optional. The skill functions without it, but users won't receive push notifications.

Conversation Examples

See references/conversation-examples.md

Error Handling

See references/error-handling.md